The person has a long recording and wants a few short pieces from it. First question: who speaks, and how long is it? One speaker under an hour is read from a transcript and the picture; several speakers need the transcript split by voice; past an hour, a model watches it first to point at the stretches worth reading closely.
Routes
| Route | Pick it when | Cost and time | What you get | Leaves |
|---|---|---|---|---|
| A. Transcript and picture | One speaker (a stream, tutorial, demo, talk), under an hour | Free; a few minutes of transcription a recording hour on the large model | A sentence menu with times, and the screens worth showing | python-contact-sheet, whisper-cli-transcribe-with-word-times, python-read-word-timings |
| B. Diarized | A podcast or interview: the moments are exchanges, and a clip shows who speaks | elevenlabs-speech-to-text at its price, with speaker ids | Speaker turns with times to cut and frame by | elevenlabs-speech-to-text, python-read-word-timings (turns) |
| C. Watch pass first | Over an hour, or the person can't say where the good parts are | gemini-script-from-video on the whole file, a few credits | A timestamped map of topics and screens to narrow route A or B to | gemini-script-from-video, then A or B on the stretches it names |
Process
- Watch it, don't just transcribe it. Contact sheets of one frame every 6 s with
the time printed, 25 a sheet, and full-size frames of the key screens
(
python-contact-sheet). Note the layout, the best visuals, anything already on screen. Do the same at 2 s for any of the person's existing clips they want matched: cut pace, caption style, title cards. - Transcribe with word times (the route's leaf) and list the sentences with their
start times (from
python-read-word-timingswords, broken at . ? ! or every ~28 words): the menu you pick from. - Pick the pieces. Each one:
- Opens on a hook a cold viewer gets: orientation first when it's needed ("what you're looking at here is my new X"), then the clever line.
- 5–9 whole sentences on one idea, ending on a payoff that doesn't repeat the hook's words. A promo bite can be one sentence; it still ends where the thought does.
- Written as sentence numbers plus the screen moment to show for each.
- Drop any piece that ends on a dangling "and…" or "but…". Check every piece's text, not just its times. Gate: the person approves the moments as text, before any cut.
- Land every edge in a pause, from the sound (
python-find-speech-pauses,snap): a piece starts in the silence before its first word and ends in the pause after its last. From word times alone: 0.06 s before the first word, 0.18 s after the last, never past the next word's start. - Check each piece by its own transcript: cut it to a WAV
(
ffmpeg-cut-audio-window) and transcribe it. It must start and end on the words picked, with nothing of the neighbours. - Level short pieces with one fixed gain each, never a loudness pass per piece
(
ffmpeg-normalize-loudness, Tips and traps).
Choosing
- Route A even for a podcast when one host carries the clip: diarization pays off when the clip is an exchange or the framing follows the speaker.
- Picture first for screen recordings. Pieces picked from the transcript alone missed the best visuals and the existing style.
- Several short bites beat one long one in a promo: four bites of 1.7–2.9 s, each ending in its pause, carried a 30 s product promo under music.
- Pick the line that says the most. A bite "people usually run out of ideas" read as a weird thing to focus on; the line that showed the product doing the work replaced it.
Avoid
- Picking from the transcript without looking at the recording.
- Cutting on word times without checking: on a long silent take they smear across the pause, and a model's own transcript times were up to 1 s off.
- An edge inside a breath or a word: snap to the pause.
- A piece that needs the sentence before it to make sense.
Related
guide-cut-best-takes: the best take of each line, when the recording is retakes rather than one long performance.workflow-board-to-summary-video(group and rank): choosing what matters from a board's stickies rather than a timeline.workflow-shorts-from-long-videoandworkflow-podcast-promo: the pieces, then the clips.
